Liberal Leaders and Liberal Success the Impact of Alternation

Home , Bloc populaire, Edward Blake, Jack Layton, John Turner, Maxime Raymond, Unity (Canada)

Liberal Leaders and Liberal Success The Impact of Alternation

Richard Johnston The University of British Columbia [email protected]

May 2018

Prepared for presentation at the Annual Meeting of the Canadian Political Science Association, University of Regina, Regina, SK, 30 May-1 June 2018.

Abstract: The Liberal party dominated 20th-century Canadian politics by dominating Quebec. Over the same period, the party regularly rotated its leadership between Quebec and elsewhere. The paper shows that leaders from Quebec were critical to Liberal success in that province, and at a minimum they did not elicit backlash in the rest of Canada. The paper reviews the history of Liberal strategy. It deploys electoral data from Confederation to the present and proposes an estimation model for parallel analyses in Quebec and the rest of Canada. Typically, impact from the Quebec leader is not instant. The paper also considers an obvious omitted variable, the macroeconomy. Unemployment is an independent factor in the Liberal vote but including it in an estimation model does not detract from the power of leadership.

Keywords: Liberal party; leadership; Alternation; Quebec; unemployment

- 2 - The Liberal party dominated 20th-century Canadian politics by dominating Quebec (Johnston 2017, Table 3.1). Over the same period, the party regularly rotated its leadership between Quebec and elsewhere. How critical were leaders from Quebec to success in that province? Did success in Quebec come at a price in the rest of Canada? Some of these questions have been taken up before, notably by Lemieux and Crête (1982), Crête and Simard (1984), and Nadeau and Blais (1993). But much has happened since those early contributions, and analysis in greater depth is now possible. It is time to return to the questions. This paper takes them up with electoral data from Confederation to the present. It begins with a review of evidence for elections and for Liberal strategy. The review is tied to recent theoretical developments in ethnic coalition politics. It proposes a simple estimation model for broadly parallel analyses in Quebec and the rest of Canada, and considers the unfolding of leadership dynamics over the full period. It then moves to a more complex model that tests economic factors as an alternative explanation for the original pattern; this too is in the spirit of Nadeau and Blais (1993). The basic pattern survives the challenge. The paper concludes by asking whether the pattern is sustainable.

Empirical and Theoretical Preliminaries A pattern for leaders from Quebec is already on the record. Lemieux and Crête (1982, 222-3) and Crête and Simard (1984, 194-6) estimate that having a leader from the province adds 13-15 percentage points to the Liberals’ Quebec vote, relative to recent levels.1 Nadeau and Blais (1993) echo this finding and extend it to the entire electorate, with a Canada-wide estimated impact between four and five percentage points. Although they conjecture (780) that the boost inside Quebec “may be amplified by bandwagon effects in the rest of the country,” they supply no proof. The direction of effect could diverge between Quebec and the rest of Canada, with the former simply overpowering the latter. The Nadeau-Blais finding is also notable in another way: leadership is not their initial concern. Rather, their focus is on economic voting, with the Quebec leadership component as a necessary control. Yet this ancillary consideration ends up dominating their system. These findings are consistent with recent work on ethnic voting patterns, indeed add up to a case in point. Ethnicity is commonly presented as susceptible to policy extremism and irreconcilable positions, in contrast to the fungibility of positions in the economic domain. A classic statement is Rabushka and Shepsle (1972). An alternative view is gaining traction, starting with Saideman et al. (2002), which emphasizes not the fixity of ethnic-group positions but their mutability. Ethnic politics as ordinary politics is the central theme of Birnir (2006). She proposes a notion of

1 The smaller number is the simple arithmetic difference as reported in Crête-Simard and the larger number is from a regression estimation in Lemieux-Crête with the current vote relative to the average in the six most recent elections. I infer that early elections in their series incorporate information from before 1930, presumably back to 1908.

- 1 - the “ethnic attractor,” which “produces a long-term pattern through the association of ethnic voters who are enticed to rely on fellow group members for electoral cues” (p. 10). The policy content of this cue-giving is commonly negotiable. Rarely, pace Rabushka and Shepsle, do demands become all-or-nothing. She goes on to observe that … [i]n the short term, cohesion dominates ethnic group behavior. In the long term, flexibility becomes the norm, as the group responds cohesively to exogenous influences (Ibid.). Cohesion and mobility lie at the very core of Johnston’s (2017) claim about Quebec as the 20th- century pivot for Canadian government. Two things are critical to making the cohesion-mobility logic work. One is that votes are cast for policy not for parties (Kedar 2009). Parties are the instruments of policy, but are not ends in themselves. Secondly, this presupposes that a party is actually able to enact policy, alone or in coalition. As Birner (2006, 38) puts it: “A party that cannot enact policy is less representative than one that can.” Either an ethnic party is mobile in choosing coalition partners or ethnic voters are mobile in choosing among parties. As credible contenders for single-party government, Canadian parties attract the second form of mobility. Quebec voters respond accordingly, or at least they did so for many decades. The complex perfectly exemplifies the pattern identified by Huber (2012): … in majoritarian systems with geographically dispersed groups, there can be strong incentives for a minority group to vote together for a mainstream party in an effort to become pivotal in determining election outcomes. (999) The other condition for this to work is that parties must be accessible to members of minority ethnic groups. Best of all, conjectures Birnir, if ethnic representation is structured through “larger, nonethnic parties” (p. 15). And access to leadership is a strong signal about access to the policy process. Evolution of the Pattern Across all elections since 1867, the Liberals were led by a Quebec francophone about half the time (see Table 1). As this was true of no elections in the first 34 years, the Quebec position for the 20th and 21st centuries is even stronger. For the last five decades, only Quebec francophones have led the Liberals to parliamentary majorities and only one never did so. In the Darwinian world of party politics, choosing such leaders evidently carries survival value. How is this value delivered? It is unlikely to be the product of dumb luck. Rather, it bespeaks intra-party elite coordination to deliver a pattern of alternation. This coordination is helped by taking the process outside the halls of parliament. This in turn required new norms, which include forbearance, and these are difficult to enforce. Ultimately, alternation begs the question, alternation of what? The next sections take these questions up in turn.

- 2 - Going Outside Alternation was helped by the shift in 1919 from intra-parliamentary selection to an extra- parliamentary convention. It mattered little that, thanks to MPs’ influence over the selection of constituency delegates, early conventions could be styled as a caucus choice in disguise (Power 1966, 371-2). The key point, rather, is that the extra-parliamentary mechanism facilitates candidacies by persons with modest experience, including from outside parliament; it expands the menu (Kenig 2009). This in turn increases the scope for managed succession (Courtney (1973, 137ff). The appeal of the outside strategy is illustrated per contra by the effort required to manage the inside transition in 1887 from Edward Blake to Wilfrid Laurier. As Blake manoeuvred to leave politics, he operated on the belief that his party’s growth prospects in Ontario were weak. He also believed that the two most obvious Ontario-based successors, Sir Richard Cartwright and David Mills, would meet resistance in Quebec; the only Ontarian acceptable to Quebeckers was Blake himself. Meanwhile, it seemed clear that in Quebec anti-Conservative sentiment was growing, in reaction to the execution of Louis Riel. Given this convergence of facts, Laurier was the solution to Blake’s retirement puzzle (Banks 1957, 119-120). Accordingly, he acted with determination to have his colleagues make the right choice (Courtney 1973, pp. 49-50). He induced Mills and Cartwright to step aside and to nominate Laurier. He acquiesced in the impression that Laurier was a temporary expedient and that Blake himself would return, a tactically useful illusion (Banks, 122-23, 125; Willison 1903, Vol. 1, 34). Of course, Laurier was far from an interim choice. Norms and their Enforcement Although Blake’s move seems motivated by the immediate electoral context, Laurier’s longevity and success may have put alternation as such on the table. At first, this worked to the advantage of non-Quebeckers. In 1919, Protestants inside the Liberal party wanted payback for loyally supporting a Catholic francophone for more than 30 years (Lederle 1947, 85-6). On the other side, “[o]ne prominent French-Canadian Liberal M. P. said … ‘We French-Canadians efface ourselves. We admit that the leader must be English Protestant …’” (quoted in Lederle, 90). When the winner at the 1919 convention, William Lyon Mackenzie King, retired in 1948, it was an “open secret” that he and his cabinet supported Louis St Laurent as his successor (Quinn 1951, 231). Their emphasis was on securing the Quebec base (Ibid., 232-33), which seemed at risk after the conscription crisis of 1944-5. But by 1948 alternation also seemed to be congealing as a norm, particularly as the Laurier-King succession had yielded stability and success (Courtney (1995, 167). If alternation was now a norm, it was not affirmed universally. In the next occasion for the choice of a francophone, 1968, a majority of delegates opposed its continuance (Regenstreif 1969). Among the reasons delegates gave for their choice, the virtue of alternation was rarely mentioned. Supporters of Pierre Trudeau, the presumptive beneficiary, were more likely than

- 3 - others to offer alternation as a reason, but (LeDuc 1971, 110-111) sees this as heavily coloured by rationalization. Trudeau’s ultimate majority required four ballots to secure and was the narrowest in the party’s history. In 1984 Trudeau himself attacked the presumption of alternation as both a barrier to talent and demeaning of the beneficiary (Chrétien 1994, 199). Given this lack of consensus, alternation requires elite initiative and timely self-restraint. When it is not the Quebec turn, potential candidates from that region must not offer themselves. This was illustrated in 1919 and again in 1957. Should a Quebecker break ranks, elites in the province must work against that person. When it is the Quebec turn, only one Quebecker should come forward and only after elite vetting. Elites outside Quebec must join those inside the province in creating a critical early bloc, making the Quebec candidate a pole of attraction for later ballots. These principles are illustrated by the contrasting cases of 1984 and 1968. In 1968, contestation was very close and the convention needed to be intensely managed. It helped that Quebec delegates were consistently the most solidary (LeDuc 1971, Tables VII and IX, 112 and 113 respectively). Even so, 40 percent of Quebec delegates did not give Trudeau as their “true” preference (ibid., Table VIII, 112). Although final-ballot differences between elite delegates and the rank and file were weak (LeDuc 1971, Table XII, 116),2 in the critical early ballots elites were distinctively consolidated around Trudeau (ibid., Table X, 115). In 1984, contrary to previous norms, Jean Chrétien sought to be Pierre Trudeau’s successor. But Quebec colleagues discouraged Chrétien and declined to endorse him: “They saw the tradition [of alternation] as a positive custom and they were reluctant to break it: one break now could open the way to a series of anglophone leaders later on” (Chrétien 1994, 194). Ambiguity of Categories The point in choosing Laurier was to carve out a place for Quebec francophones. But this begged the question of which was more critical, language or region (or, for that matter, religion). The ambiguity is perfectly captured by Paul Martin, Sr and Paul Martin, Jr. The elder Martin’s background was bilingual but situated mainly on the French side of the divide. He emerged from the 1917 conscription crisis as quite nationalist (Donaghy 2015, 8 et supra). When he ran against Lester Pearson to succeed Louis St Laurent in 1958, he was handicapped by being too French and too Catholic (Donaghy, 162). In 1968, he was not French enough (Martin 1983): he had chosen the anglophone path for university education and the bulk of his public life was solidly rooted in Ontario. Paul Martin, Jr presents himself as an anglophone and was born and educated in Ontario. But his business and political life lies squarely in Quebec. His defeat by Jean Chrétien in 1990 and his accession to the leadership in 2003 can be characterized only as preserving alternation if we stipulate that Quebec leader must also claim native fluency in French.

2 Throughout, the most distinct delegates were Young Liberals, tilted toward John Turner (ibid., 115-116).

- 4 - This logic extends to the succession in 2006. Stéphane Dion was the sole francophone in a race that was arguably the francophone turn. But his victory also implied unbroken leadership from Quebec. In the early stage of the leadership campaign, this worked against him: in an editorial that ultimately endorsed Dion, the Toronto Globe and Mail (2006) stated, “[a] sense that it is the turn of a leader from outside Quebec has worked against him.” The editorial said nothing about linguistic alternation, nor did it endorse him on that explicitly categoric basis. The next successions—Michael Ignatieff, anglophone from Ontario, and Justin Trudeau, from Quebec—might seem to clarify matters. But Justin Trudeau is less essentially French and less rooted in the politics of his home province than any Quebec-francophone predecessor. And the high frequency of leadership change—four switches in less than a decade—arguably undermined the deliberative and controlled character of the earlier, slow-moving pattern, where a leader typically held the position for at least ten years or was allowed to weather at least one defeat.

Indicators, Estimation Strategy, and Interpretation Leader Indicators Full representation of parties’ leadership requires three indicators for Quebec and two for the rest of Canada. Indicators for the Liberal and Conservative parties appear for both sets of estimations. The reason these two parties’ leaders appear in both estimations should be obvious, the parties run candidates everywhere. The significance of the leader’s province of residence may matter more to Quebeckers than to others but there is no reason to dismiss by fiat the relevance of the home province to both electorates. The third indicator is for parties that I term “Quebec ethnonational,” and here the relevance is Quebec-specific. The appearance of the leader of such a party bespeaks the party’s very existence. This in turn suggests some sort of crisis that disrupts the historic domination of the province by the two old parties. Additionally, the simple arithmetic of party competition suggests that as the menu expands in Quebec, the space left over for the old parties shrinks. This arithmetic does not apply where Quebec ethnonational parties do not run candidates, that is, outside Quebec. The three parties/party types appear so as to mirror the logic in Nadeau and Blais (1993). But as Nadeau and Blais had a small number of observations, they conserved degrees of freedom by loading the three party situations into a single indicator. This forced them to commit ex ante to arbitrary values for each component even as the avenues of effect are masked. My setup allows each leader type to have its own effect. Coding decisions for each category require justification, however. For the Conservatives, the big question is Brian Mulroney. He is not a native francophone but his fluency and command of idiom are complete. Moreover, his initial entourage tended to identify quite closely with Quebec national aspirations (Sawatsky 1991). For Nadeau and Blais (1993), Mulroney qualifies as a Quebec leader fully in the class of his Liberal counterparts; I follow their example. The 1997 election also meets this test, with Jean Charest at the helm. Even so, a Quebec-based Conservative leader is a rare thing, present in only 7 percent of all elections.

- 5 -

Table 1. Descriptives (N = 42)

Variable Mean SD Min Max

Liberal Vote

Quebec 45.9 14.7 14.2 73.4

Rest of Canada 37.8 7.2 21.7 50.7

Leader from Quebec

Liberal 0.52 0.51 0 1

Conservative 0.07 0.26 0 1

Quebec ethnonational 0.36 0.48 0 1

Unemployment

Quebec 8.2 2.52 3.4 13.3

Canada 6.6 2.23 2.8 11.5

The Quebec ethnonational category also requires exercises of judgment. The earliest election with such a leader is 1945, in the aftermath of a conscription crisis. The primary entity was the Bloc Populaire Canadien3 led by Maxime Raymond, but the landscape also featured an uncoordinated array of Quebec nationalist insurgents (Cohen 1965, Table 7; Johnston 2017, 94). Also included are all elections in which some variant of Social Credit was led by a Quebecker, 1965 to 1980 inclusive. In 1962 and 1963, the Quebec contingent massively outnumbered the remnant from the rest of the country, but the nominal leader of the party was Robert Thompson from Alberta. For some of this period the Ralliement Créditiste was a separate organization from the ostensibly pan-Canadian Social Credit. When the two Social Credit rivals reunited, the Quebec wing was the senior partner. Throughout, the Créditistes maintained a loose social- conservative and nationalist affinity with the remnant of the old Union Nationale. All elections

3 As the name suggests, for this Bloc the nation was French Canada, an entity both more and less inclusive than Quebec. Even so, the party ran only two candidates outside the province and there is no reason to treat this grouping as anything other than Quebec-specific.

- 6 - from 1993 forward fit the ethnonational criterion, with a Bloc Québécois leader continuously present. All told, such parties have been present in 36 percent of all elections. Finally, there is the leadership of the Liberal party. At the time Crête-Lemieux-Simard and Nadeau-Blais were writing there had never been a Quebec leader who was not a francophone nor a francophone who was not a Quebecker. Since then, as the earlier narrative showed, the party has been led by Paul Martin, who presents as essentially anglophone, and Justin Trudeau, an ambiguous case. Nadeau and Blais show the way, I believe, with their treatment of Brian Mulroney: the key is that the leader represents a riding in Quebec; the rest follows from that. Thus, the chain of Quebec leadership, as coded, is unbroken from Chrétien to Dion. Overall, slightly more than half of all elections feature a Liberal leader from Quebec.4 Economic Indicators The second part of this paper deploys economic evidence. Apart from the general interest in updating the Nadeau-Blais analysis, there is also a question of substantive interpretation of leadership effects. Brian Mulroney’s accession to the Conservative leadership in 1983 coincided with the highest level of unemployment in the entire postwar period, with 1984 close behind. His 1993 departure coincided with another deep recession and the second highest postwar unemployment reading. The 1962 election, the breakthrough year for the Ralliement Créditiste, followed the so-called Eisenhower recession. The year before this election saw the third highest postwar unemployment rate and the highest to that date. (Recall, however, that Réal Caouette was not inscribed as the Créditiste leader until 1965.) The unemployment peak of 1993 was also the year the Bloc Québécois fought its first election. It is natural to wonder if the apparent effects from Conservative and Quebec-ethnonational leaders are not so much from the persons as from the times. Following Nadeau-Blais, I tested models with inflation, real income growth, and unemployment. Also as in Nadeau-Blais, inflation and real income growth had no effect. Unemployment is another matter. I use the average level by year, from seasonally-adjusted source data. In Quebec, election-year unemployment has ranged almost 10 points. The range in Canada as a whole is about one point smaller. Over the full period, unemployment in Quebec has stood 1.6 points higher than the Canada-wide average. Despite these discrepancies, the time path is basically the same in each part of the country.

4 This criterion also excludes leaders of any party who did not represent a Quebec riding at the time of leadership. For the Liberals, this excludes John Turner, whose first seat was in Quebec. For the NDP, this excludes David Lewis, who spent most of his childhood in Quebec and graduated from McGill but whose electoral career was entirely based in Toronto. Likewise, it excludes Jack Layton, who was born and raised in Quebec and was also a graduate from McGill. As with Lewis, his political base was Toronto. In 2015, the NDP leader was from Quebec. It seems ill advised to create an NDP leader variable, given that this happened only once and only at the very end of the series.

- 7 - My estimation model differs in a critical particular from the Nadeau-Blais one. Their focus is on the incumbent, such that the effect of unemployment is exactly reversed depending on which party is the incumbent. My setup is open to the possibility that some, at least, of the impact from unemployment is positional. That is, voters react not just to whether or not a party is in power but also to what they see as its policy priorities, a possibility first proposed in political science by Kiewiet (1981). In recent work that is both synoptic and attentive to mechanisms, Kayser and Grafström (2016) show that the negative impact of unemployment is particularly harsh for incumbents on the left, and that this is so across a wide range of countries and continents.5 Moreover, the qualitative pattern may extend to periods when the left is out of power. On this logic, the Liberals may be generally susceptible to unemployment effects although especially so when they are in power. But are Canadian Liberals on the left? There is little indication that Liberal governments produce better redistributive outcomes than the Conservatives or that they are more concerned than the Conservatives to fight unemployment. The Kayser-Grafström argument points us away from such considerations, however. To them, the key is that the left is identified with the provision of “luxury goods,” policy outcomes that may be valued when times are good but that drop in priority when times turn bad. These include human rights, multiculturalism, and environmental protection. On these issues, the Canadian Liberals stand clearly to the left of the Conservatives. With this in mind, incumbency is represented in the estimation by a simple dummy. This imposes no restraint on how, if at all, incumbency conditions the effect of unemployment; it neither precludes symmetry nor requires it. Observations Lemieux-Crête-Simard cover fifteen elections in five decades, 1930-1979. Nadeau and Blais cover 11 or 13 elections, 1953-88.6 It is straightforward to update the analysis as we have seen eight elections since Nadeau-Blais and eleven since Lemieux-Crête-Simard. For analyses that exclude economic factors there is no compelling reason to start in 1930. Extending the series enables inclusion of the breakthrough case of Sir Wilfrid Laurier and expands the number of observations to 42. It is true that in the early years the system underwent institutional and demographic change. The parties required some decades after 1867 to fully colonize the electoral landscape. Only in 1874 were writs of election issued simultaneously for all ridings, and only later for ridings in the West (Reid 1932). The secret ballot was introduced in 1878 (Ibid.). The franchise remained restricted in the late 19th century and although it came to comprise nearly all male citizens in the early 20th century (Ward 1950), women did not enter the electorate until

5 Their finding is echoed in Dassonneville and Lewis-Beck (2013), even though the focus for the latter is subtly different 6 They treat 1958 and 1980 as borderline cases for the retrospective-voting logic of their model, in that the Conservative incumbents in each year had been in power for less than a year and so hardly qualify as responsible for the current state of the economy. Accordingly, they estimate the model twice, once with and once without these elections.

- 8 - 1917 and 1921. Initiating the series at later dates to accommodate these developments does not alter the basic structure of the analysis, however. As mentioned, the only relevant economic series is for unemployment. As this series starts in 1946, it yields 22 observations with the 1949 election as the first. A critical element for all estimations is that the vote series be stationary, which is to say that it is “stable” and that it exhibits minimal trending, periodicity, or structural breaks. Both vote series outside Quebec pass these tests with flying colours. In Quebec, the conditions for estimation are satisfied for the 1867-2015 series. For the postwar series, the situation is not so clear. Although the resultant estimation must be treated with some scepticism, its structural similarity to the longer-period estimation is reassuring. The relevant diagnostics appear in the Appendix, Section 1. For estimation by OLS, the concern will be that disturbances are serially correlated. I return to this matter in discussing the findings.

Estimation Strategy The basic strategy is to regress the Liberal vote on dummy variables indicating whether or not the leader of the designated party is from Quebec. My setup is open to the possibility that the electorate does not respond instantly to the change of leadership. Start by assuming per contra that the electorate adjusts to an exogenous change instantly:

∗ (1) � = � + �� + � , � = 1, … , � ∗ where � is the hypothetical Liberal share at election t depending on whether or not � = 1, the leader is a Quebec francophone.7 To the extent that there is inertia in the system, actual shifts over an interval, will be expressed thus:

∗ (2) � − � = � (� − � )

� is the observed Liberal share at t and λ indexes how much of the inertia-free potential is actually realized over the interval, where 0 < λ ≤ 1. If λ = 1, then the shift is instantaneous. ∗ Although � is not observed, we can model it by substituting (1) into (2):

� − � = � � + �� + � − � .

Exploding the parentheses, rearranging terms, and factoring � yields:

(3) � = �� + �� + 1 − � � + �� , which now looks like something we can estimate. If we set �� = � , �� = � , 1 − � = � , and �� = � , we can estimate the equation as:

7 For simplicity of exposition, I present only one leader term, even though the actual estimation deploys three for Quebec and two for the rest of Canada, one for each relevant party or party type. The exposition also applies to impact from shifts in the unemployment rate.

- 9 - (4) � = � + � � + �� + � . The most important interpretive discrepancy between the actual regression estimates in (4) and the theoretical quantities in (3) pertains to �, the estimated marginal effect of a switch to a Quebec francophone leader. That estimated effect is no longer the full effect that we suppose might occur in a world of instantaneous adjustment, that is, �. Rather, � = ��, a version discounted by the speed of adjustment. To recover the full effect of an indefinitely sustained shift to a Quebec francophone, we need to adjust the regression estimate, thus: � � , the sum of an infinite series. But λ is also a derived value. The derivation starts with δ, the coefficient on the lagged dependent variable. Given that δ = 1 - λ, it must be the case that, λ = 1 - δ. Hence, � = � (1 − �). The presence of the lag term reflects the fact that gradual onset is a realistic possibility. For instance, Evans and Andersen (2006) show that past vote affects current economic perceptions and the same is likely to be true for leader perceptions. For reasons like this, political time series are commonly modelled as systems of such partial adjustment.8 Additionally, my strategy is a variant of the “general-to specific” approach (Hendry 2003), which argues that there is little justification for imposing instantaneous adjustment by fiat, as would be the case without a lag term in the model. Not incidentally, the lag term can also help soak up autocorrelated in errors (Pickup 2014), such that estimation by OLS is possible. (In fact, for precisely this last reason, we always need two lag terms in estimations for the rest of Canada.) The basic expectations are:

• � > 0 (or � > 0). When the Liberals choose a leader from Quebec their vote in the province increases. • For leaders of the Conservatives and of Quebec-based ethnonational parties,

� �� < 0. This logic applies to Quebec. For the rest of Canada, the signs of effect for the Liberals and Conservatives could be the same as in Quebec or the opposite. There is no reason to expect any effect at all from the presence or absence of Quebec-based ethnonational party. Accordingly, the term does not appear in the non-Quebec estimation. Given that the non-Quebec estimation requires two lags, it seems better not to burn up this additional degree of freedom. Whatever the sign of the relationships outside Quebec, the strategic logic of alternation applies to Quebec itself. Hence, I expect the following to be true:

8 The alternative to partial adjustment is a system that involves error correction. This would involve a slightly different and less readily interpretable setup. For early discussions of the merits of the alternatives, see Beck (1991). Grant and Lebo (2016) offer a spirited case that error correction is inappropriate for all but a small fraction of political data.

- 10 - • | �| > | �| .

Basic Estimations Quebec The evidence for both Quebec and the rest of Canada appears in Table 2. All estimations are by Ordinary Least Squares (OLS). See the Appendix, section 2 for details on the choice of estimation strategy. According to Table 2, choosing a leader from Quebec makes a huge difference to the Liberals’ fate in that province. For the Liberals, the immediate gain is about eight percentage points. When the Conservatives choose such a leader, the hit on the Liberals is even bigger, about eleven points. A third party with a Quebec leader and a primarily or exclusively Quebec base reduces the Liberal share by about half that amount (although the coefficient is not unequivocally different from zero). The immediate effect is not the whole story, however. The lag term, denoted as Vote (t-1), implies that only one-third of the total potential gain occurs on the first iteration. Should a Quebec leader remain in place for, say, five or six elections, the total gain would be greater than twenty points. Only two leaders come close to such durability, of course, and the estimates in Table 2 are powerfully affected by the longevity of Wilfrid Laurier and Pierre Trudeau. To get a sense of how well the model fits the facts, consider Figure 1. The figure embodies predictions from Table 2, equation 1 (dashed grey), along with the observed values (solid black). Bear in mind that the model has other moving parts: three observations with a Conservative leader from Quebec; several observations with a Quebec third-party; and movement in the lag term not just from leader factors but also from forces outside the model. Wilfrid Laurier and Pierre Trudeau illustrate Quebec leadership over a long haul. In Laurier’s first election as leader he boosted the party’s fortunes not a bit. The previous election, in 1887, was the breakthrough, a doubling of the party’s share in Quebec. It may be significant that Laurier’s first platform, in 1891, included “Unrestricted Reciprocity,” a proposal to drop tariff barriers on US goods and ask nothing in return. This was not popular with the Montreal business community. Added value from the switch to Laurier came only in 1896, when reality almost caught up with prediction. This was in spite of the fact that the Catholic hierarchy endorsed the Conservative position on the Manitoba Schools Crisis, which was then unfolding. Arguably, voters saw Laurier and his endorsement of robust federalism as the better guarantee of communal rights (Crunican 1974). The subsequent time path of predicted values roughly tracks that for observed ones—until 1911. The 1911 election revealed the continued existence of anti-Rouge and ultra-nationalist forces in Quebec, which produced a negative reaction to the 1910 Naval Bill (Brown 1975, 163ff, 235 passim). The observed drop in 1911 generates a further predicted drop in 1917. Instead, the conscription crisis sent the Quebec Liberal vote through the roof. Critically,

- 11 - Laurier was intransigent on conscription; a leader from another province might have handled the

Table 2: Quebec Leaders and the Liberal vote

(1) (2) Quebec Rest of Canada N = 41 N = 40

Coeff SE(b) 95% CI Coeff SE(b) 95% CI

Leader from Quebec (t)

Liberal 7.99 3.22 1.46, 14.52 1. 52 2.16 -2.87, 5.91

[Long-run] 23.21 11.29 1.08, 45.35 2.02 2.74 -3.35, 7.39

Conservative -11.29 5.64 -22.72, 0.14 -5.48 -3.85 -13.32, 2.35

Third-party -5.70 3.45 -12.70, 1.30

Vote (t – 1) 0.66 0.10 0.44, 0.87 0.46 0.17 0.12, 0.81

Vote (t - 2) -0.21 0.17 -0.56, 0.13

Intercept 14.77 5.54 3.52, 26.01 35.38 7.30 20.55, 50.21

!" 0.63 0.25

RMSE 9.33 6.00

Estimation by OLS. Long-run statistics estimated with nlcom in Stata.

matter differently. As with Laurier, Trudeau’s relationship to his province was complicated. His accession coincides with a jump in Liberal share, although smaller than predicted. The next two elections are broadly in range with prediction, although with short-term motion contrary to expectation. The backdrop to 1972 includes the 1970 October crisis, with two kidnappings, a political murder, and the abridgement of civil liberties (Smith 1971). The 1974 election shifted focus to the economy and restored the Liberals’ share. In Trudeau’s last two elections, the Liberal share reached heights unheard of since the 1917 conscription crisis. This was a period of mounting constitutional

- 12 - tension, with Quebec as the focus. Also as with Laurier, the largest pro-Liberal shift occurs not upon accession but at a later moment of crisis.

Figure 1. Liberal leader dynamics in Quebec

Notes: Entries are Liberal shares of the Quebec popular vote. Markers indicate elections with a Quebec francophone leader. Quebec-based Conservative and third-party leaders are also included in the estimation. Predicted values from Koyck lag estimation in Table 2, equation (1).

The other two Quebec leaders to last more than one election follow contrasting paths. St Laurent was an instant hit and recorded no further gains. The prediction model says that the Liberal share should grow again in 1953, but it did not. All three readings for St Laurent are higher than all but one recorded by Laurier. Chrétien followed the classic path, if on a relatively low absolute level. There is an uptick in his first election, 1993, but less than predicted. The 1997 election was supposed to yield a poor result, thanks to Jean Charest assuming the Conservative leadership. Instead, the Liberals made a modest gain. Their 2000 result was exactly as forecast; the forecast incorporated both the gain in 1997 and the departure of Charest from federal politics.

- 13 - The two leaders with one election to their name, Stéphane Dion and Justin Trudeau, both seem to have helped their party, Trudeau especially. The era for both is a bad one for the Liberals, however. The 2008 result is the worst since 1887 and the 2015 one, even though it yielded a handsome crop of seats, is the third lowest. Although observed values do not march in lockstep with predicted ones, the patterns are broadly consistent with the logic of the partial-adjustment setup. With the sole exception of Louis St Laurent, the first election with a Quebec leader yields modest results. The two longest-serving ones ended their careers on much higher notes than they started them. The same is true of Jean Chrétien. Only for St Laurent was the last result the worst. Dion and Trudeau, of course, have no trajectory to discuss.9 The Rest of Canada In the rest of the country, the pattern is a pale echo of that in Quebec. For neither party is the effect significantly different from zero, but the Conservative finding is more convincing than the Liberal one. The long-run effect is similarly pallid. Not only does the first lag hold back only half the effect of the total shift but the second lag partly offsets the first. Thus, the initial effect and the long-run effect are barely distinguishable. Accordingly, the dynamics outside Quebec are less remarkable overall than in Quebec and less readily tied to leadership. The three biggest drops occurred with a Quebecker at the helm: 1896 and 1917 under Laurier and 1957 under St. Laurent. But none of these drops was forecast in any sense and the party did just fine under those leaders most other years. The drop in the early 2000s is also striking and occurred under Quebec leaders, but again this followed a decade of quite satisfactory outcomes.

9 The plot in Figure 1 has worrisome extreme outliers. For an account of their impact on the overall estimation, see the Appendix, section 3. That section also discusses outliers for the rest of Canada estimation.

- 14 - Reversion to a non-Quebec leader is not a magic bullet. It is true that the Liberal share went up when Mackenzie King succeeded Laurier. But the initial gain was modest, and probably reflected more the catastrophe of 1917 than the peculiar appeal of King. King’s striking gains came later and hardly seem attributable to shedding the Quebec burden. The succession after St Laurent was similarly unimpressive. Lester Pearson’s loss in 1958 was almost as bad as that in 1917. The recovery in the early 60s was tepid. Pierre Trudeau clearly dissipated the credit he gained in 1968 but the disaster of 1984 came on John Turner’s watch, and 1988 was scarcely better. Arguably, the best way to repair the damage left by one Quebec leader is to replace him— eventually—with another one.

Figure 2. Liberal leader dynamics outside Quebec

Notes: Entries are Liberal shares of the popular vote outside Quebec. Markers indicate elections with a Quebec francophone leader. Quebec-based Conservative and third-party leaders are also included in the estimation. Predicted values from Koyck lag estimation in Table 2, equation (2).

- 15 - The Person or the Times? The rationales for extending the analysis to include the economy, for the indicator, and for the conditional setup have already been given. The presentation is in two stages, a baseline for calibration relative to the full-period estimations in Table 2, and the unemployment estimation itself.

Table 3. A Model with Unemployment, Quebec, N = 22

Baseline + Unemployment

Coeff SE Coeff SE

Leader from Quebec (t) Liberal 11.41 5.06 12.67 5.77

Conservative -8.90 6.05 4.67 9.19

Third-party -4.49 5.38 4.26 6.96

Vote (t – 1) 0.66 0.16 0.90 0.22

Incumbent -0.76 19.71

Unemployment -1.60 1.95

Incumbent x Unemployment -1.02 2.07

Combined effect -2.62 1.44

Intercept 10.43 9.01 9.95 15.59

!" 0.59 0.60 RMSE 9.23 9.03

Note: The unemployment series is from Statistics Canada Table 282-0087. Rate is for Quebec.

The Quebec pattern appears in Table 3. The baseline estimation is quite close to that in Table 2. The Liberal coefficient looks stronger, the Conservative coefficient weaker, and the Quebec ethnonational coefficient also slightly weakened. Only the Liberal coefficient is significantly different from zero. The pattern is basically recognizable, if tilted toward leadership choices by the Liberals. Tests for equivalence of coefficients in the full and truncated series indicate that

- 16 - they are essentially equal; for this see the Appendix, section 4. With this we can move to the estimation that includes the unemployment effect. In the full model, the relative ranking of leader effects remains unchanged but only the Liberal one remains basically as before. Indeed, it seems strengthened. The other leader coefficients turn into noise: their signs reverse—a nonsense result—and their standard errors expand. The economy itself does seem to matter and in a differential way. Even when they are out of power, the Liberals are affected adversely by unemployment. This pattern may help account for the Liberals’ weak performance in Quebec as they moved to regain power in 1962-3 and 1993. According to Figure 1, their 1962 share was the lowest to date since 1887 and their 1963 share,

Table 4. A Model with Unemployment, Rest of Canada, N = 22

Baseline + Unemployment

Coeff SE Coeff SE

Leader from Quebec (t) Liberal 4.64 3.01 4.83 2.97

Conservative -2.92 3.92 3.44 4.37

Vote (t – 1) 0.24 0.25 0.17 0.30

Vote (t – 2) -0.23 0.26 -0.49 0.27

Incumbent 15.33 10.26

Unemployment -0.08 1.18

Incumbent x Unemployment -2.24 1.32

Combined effect -2.32 0.84

Intercept 33.08 9.80 43.67 16.61

!" 0.09 0.30 RMSE 5.83 5.14

Note: The unemployment series is from Statistics Canada Table 282-0087. Rate is for Canada.

the lowest (1962 aside) since 1911. But when the Liberals do attain power, things get worse: an

- 17 - additional percentage point is added to the impact estimation. Taken separately, neither coefficient is significantly different from zero. More striking, however, is the estimate for the combined effect of incumbency and unemployment, the numbers in bold. Here the coefficient is, as we could already calculate, about 2.5 points and it is almost twice its standard error. It has a 0.09 chance of being the product of chance covariance. The same patterns appear outside Quebec, according to Table 4. In the baseline estimation, signs on the leader variables are as in Table 2, although now with the Liberal leadership more clearly dominant than before. As with the Quebec estimations, the coefficients are officially indistinguishable from those for the full period; for confirmation, see the Appendix, section 4. Adding the incumbency and unemployment terms leaves the Liberal effect in place but upsets the Conservative one, also an echo of the Quebec pattern. Although the economic pattern seems pretty clearly to be an incumbent-oriented one, it is still positional. Out of power, the Liberals are unaffected by unemployment; in power, they are clearly punished for it. The coefficient is not quite as larger as in Quebec but its estimate is clearly more stable. The probability that the combined effect is the product of chance covariance is less than 0.02. There is more than a suggestion, then, that some of the apparent leader effects in the full estimations are byproducts of the times. Postwar activation of nationalist opinion in Quebec was most intense in recessions. This was a theme in Pinard (1966; 1975) for the Créditistes and should be one for the Bloc Québécois, as Johnston et al. (1996) suggests. The fact that the Conservatives were faced with a leadership crisis in 1983 reflected concern that the party could be better poised to exploit the opportunity presented by that year’s recession. Mulroney’s decision to step down in 1993 was clearly influenced by the low state of the party’s prospects. Put this way, the leadership options that emerged for voters —for Quebec voters especially— may have been a mechanism more than a primum mobile, a transmission belt for economic forces. The same cannot be said for leader selection by the Liberals. The only transition with remotely plausible economic content was that in 1984, in the wake of a deep economic crisis. By this time, however, Pierre Trudeau’s run of power spanned sixteen years with only one brief glitch. Retirement seemed like a natural step. The other changes followed an internal organizational rhythm. St. Laurent had been groomed for 1949. His retirement in 1957 followed electoral defeat. Although there were anticipatory signs of the recession of 1958, the economy generally seemed strong in 1957. It also seemed strong in 1968, so much so that the chief concern of the time was the early indications of the inflation that came to dominate the 1970s. Jean Chrétien’s timing was fortuitous but dictated mostly by the fact that John Turner had lost two consecutive elections. His displacement by Paul Martin, Jr, was fraught but the infighting that accompanied it was of long standing. It is hard to see the economy at all in the rapid turnover that followed Martin’s time at the helm. And, to go back to the evidence in Table 4, adding the economic dimension leaves the basic pattern intact for Liberal leadership: having a leader from Quebec matters a lot inside that province, an effect that usually grows with time; outside Quebec the

- 18 - pattern is less clear but there is no suggestion that a Quebecker hurts the party and a strong hint that it helps. The economic estimations are more suggestive than definitive, however. They require slicing the data very thinly: in each full estimation, only fourteen degrees of freedom remain for the residuals. As the Appendix indicates, serial correlation in errors is problematic. In Quebec, the baseline estimation is basically free of such correlation but adding the economic terms moves the estimation into a danger zone for both first- and higher-order autocorrelation. In the rest of Canada, the pattern is the reverse. Without the economic terms, the estimation appears to harbour first-order autocorrelation. Including the economic terms seems to absorb the autocorrelation, such that the OLS estimation is unbiased.10

Conclusion and Discussion When the Liberal party chooses a leader from Quebec its vote in the province increases significantly. It is not penalized for this—and may even be helped—in the rest of Canada. On the Quebec side, the pattern helps explain the province’s role as the 20th century pivot for governments (Johnston 2017). Although the pattern outside Quebec is weaker, the sheer number of voters makes the rest of Canada essential to the overall pattern. At a minimum, voters outside Quebec must not counter mobilize. The relative sizes of the electorate, historically roughly 3:1, are such that even a modest adverse reaction outside Quebec could neutralize shifts inside the province. If estimates in Table 4 are credible, absolute gains and losses in the rest of Canada roughly match those inside Quebec. And the pattern may be self-fulfilling: since 1960 the party has been led by a Quebecker more often than not. Indeed, dropping the Quebec leader creates a moment of peril. The essence of the story survives the most obvious omitted-variable challenge, from the macroeconomy. It turns out the economy matters, roughly on the scale and along the policy channel that Nadeau and Blais (1993) identified. The exact structure of effect is not the simple retrospective one that they posited, although it clearly has a retrospective component. The component just happens to be conditional on party positioning. Although invoking the economy seems to explain away most of the apparent effect from leaders of parties other than the Liberals, it touches the heart of this paper’s claim not at all. If anything, bringing the economy into the picture sharpens the relief for Liberal leaders. The paper is largely silent on the mechanisms that produce the main pattern. This is true for both Quebec and the rest of the country. The time path of effect, both as modelled by the partial adjustment framework and as revealed by Figure 1, point away from a simple “favourite-son” type of effect. Only once was the full effect delivered immediately upon accession. This was in

10 It is reasonable to ask of all the estimations in this paper if lagging the dependent variable, as opposed to lagging the independent variables, is the correct representation of history. The answer is yes. For a discussion see the Appendix, section 5.

- 19 - 1949 with Louis St Laurent. It might have been mere reversion to a normally high level, a reflection of the Liberal party’s near-hegemonic domination of Canadian politics in the period. The reversion would be from the crisis moment of 1945, the immediate aftermath of struggle over conscription. What is more, the paper is silent on who within Quebec delivers the bump. In recent years, this may well be anglophones and allophones, rallying to a francophone federalist—and falling away when that prospect is no longer dangled. This seems unlikely for earlier years, however. Whatever the pathway, it remains to be explored. Mechanisms in the rest of Canada are similarly opaque. Nadeau and Blais (1993) refer to “strategic voting.” On this logic, voters outside Quebec register Liberal strength inside Quebec and react accordingly. Since the 1940s, they have had access to poll information. In earlier decades, they would have to base inference on the mere identity of the Liberal leader.11 Another possibility is that the Liberal party benefits from a “national unity” vote, where the choice of a Quebecker reaffirms the party’s ability to manage the issue. Yet another is that leaders from Quebec are, or have become, more competent than their non-Quebec rivals. Their very longevity may attest to this. The difficulty with this argument, of course, is its circularity. For the Liberals at least, the circle is a virtuous one, not one that harms the party for the long run. The pattern may now be living on borrowed time. Successful, enduring party building in Canada has historically required brokerage. Brokerage is helped by deliberate, multi-decade cycles that enable inter-communal trust to flourish. Long cycles also enable successors to be identified and groomed. Cycles seem to be shortening, even as the very boundary criteria for alternation are being challenged. That said, a countertrend may be underway, a trend that reflects the logic identified by Birnir (2006). At the core of her argument is that mere representation is a weak claim for an ethnic party. Substantive impact on policy is the ultimate trump, and in a Westminster-style system this means a large poly-ethnic big tent. The last two elections witnessed the collapse of the Bloc Québécois vote. So long as the Bloc had a strong foothold, it limited the ability of Quebec to serve as the veto player. Although the positive marginal impact of choosing a Quebecker to lead the Liberal party persisted, it operated against a moving baseline that was increasingly distant from the target. If the new trend continues, the old pattern may be restored. Whether that also delivers a lengthened time horizon for the Liberals is another question.

11 This is not quite the same as the argument in Johnston (2017). He sees Liberal strength in Quebec as creating a permanent strategic advantage for the Liberals, to the particular detriment of the NDP. This advantage also has implications for the relative power of cultural, as opposed to economic, issues. But his conception of the advantage is mainly static, not sensitive to the presence or absence of a Quebec-based leader.

- 20 - References Banks, Margaret A. 1957. The Change in Liberal Party Leadership, 1877. Canadian Historical Review 38 (2): 109-28. Beck, Nathaniel. 1991. Comparing Dynamic Specifications: The Case of Presidential Approval. Political Analysis 3: 51-87. Belsley, D. A., E. Kuh, and R. E. Welsch. 1980. Regression Diagnostics: Identifying Influential Data and Sources of Collinearity. New York: Wiley. Birnir, Jóhanna Kristín. 2006. Ethnicity and Electoral Politics. Cambridge, UK: Cambridge University Press. Brown, Robert Craig. 1975. Robert Laird Borden, A Biography. Volume I: 1854–1914. Toronto: Macmillan. Chrétien, Jean. 1994. Straight from the Heart, 2nd edition. Toronto: Key Porter Books. Courtney, John C. 1973. The Selection of National Party Leaders in Canada. Toronto: Macmillan. Courtney, John C. 1995. Do Conventions Matter? Choosing National Party Leaders in Canada. Montreal and Kingston: McGill-Queen’s University Press. Crete, Jean, and Johanne Simard. 1984. Conjuncture économique et elections: Une étude des élections au Québec, in Jean Crête, ed., Comportment electoral au Québec. Chicoutimi: Gaetan Morin, pp. 165-97. Crunican, Paul. 1974. Priests and Politicians: Manitoba Schools and the Election of 1896. Toronto: University of Toronto Press. Dassonneville, Ruth, and Lewis-Beck, Michael. 2013. Economic Policy Voting and Incumbency: Unemployment in Western Europe. Political Science Research and Methods 1(1): 53-66. Donaghy, Greg. 2015. Grit: The Life and Politics of Paul Martin Sr. Vancouver: UBC Press. Globe and Mail [Toronto]. 2006. The Liberal leadership: And then there was one. [25 November 2006] accessed at https://beta.theglobeandmail.com/opinion/the-liberal-leadership-and-then- there-was-one/article1329611/?ref=http://www.theglobeandmail.com& Grant, Taylor, and Matthew J. Lebo. 2016. Error Correction Methods in Political Time Series. Political Analysis 24: 3-30. Hendry, D. F. 2003. Dynamic Econometrics. Oxford, UK: Oxford University Press. Huber, John D. 2012. Measuring Ethnic Voting: Do Proportional Electoral Laws Politicize Ethnicity? American Journal of Political Science 56: 986-1001. Johnston, Richard. 2017. The Canadian Party System: An Analytic History. Vancouver, BC: UBC Press.

- 21 - Kayser, Mark A., and Cassandra Graftström. 2016. The Luxury Goods Vote: Why Left Governments are Punished More for Economic Downturns. Berlin: Hertie School of Governance, unpublished manuscript. DOI: 10.13140/RG.2.2.24409.31841 Kedar, Orit. 2009. Voting for Policy, Not Parties: How Voters Compensate for Power Sharing. Cambridge, UK: Cambridge University Press. Kenig, Ofer. 2009. Democratization of party leadership selection: Do wider selectorates produce more competitive contests? Electoral Studies 28 (2): 240-7. Kiewiet, Roderick D. 1981. Policy-Oriented Voting in Response to Economic Issues. American Political Science Review 75:448–59. Lederle, John W. 1947. The Liberal Convention of 1919 and the Selection of MacKenzie [sic] King. Dalhousie Review 27 (1): 85-92. LeDuc, Lawrence. 1971. Party Decision-Making: Some Empirical Observations on the Leadership Selection Process. Canadian Journal of Political Science 4 (1): 97-118. Lemieux, Vincent, and Jean Crête. 1982. Quebec, in Howard R. Penniman, ed. Canada at the Polls, 1979 and 1980. Washington, DC: AEI, 208-25. Martin, Paul. 1983. A Very Public Life: So Many Worlds. (Volume 2) Ottawa: Deneau. Nadeau, Richard, and André Blais. 1993. Explaining Election Outcomes in Canada: Economy and Politics. Canadian Journal of Political Science 26 (4): 775-90. Perron, Pierre. 2006. Dealing with Structural Breaks, in Palgrave Handbook of Econometrics, Vol. 1: Econometric Theory, K. Patterson and T.C. Mills, eds., Basingstoke UK: Palgrave Macmillan, 278-352 Pickup, Mark. 2014. Introduction to Time Series. Thousand Oaks, California: Sage. Pinard, Maurice. 1966. La faiblesse des Conservateurs et la montée du Crédit social en 1962. Recherches Sociographiques 7: 360-3. Pinard, Maurice. 1975. The Rise of a Third Party: A Study in Crisis Politics. Montreal: McGill- Queen's University Press. Power, Charles Gavan. 1966. A Party Politician: The Memoirs of Chubby Power, edited by Norman Ward. Toronto: University of Toronto Press. Quinn, Herbert. 1951. The Third National Convention of the Liberal Party. Canadian Journal of Economics and Political Science 17 (2): 228-33. Rabushka, Alvin, and Kenneth A. Shepsle. 1972. Politics in Plural Societies: A Theory of Democratic Instability. Columbus, Ohio: Merrill. Regenstreif, Peter. 1969. Note on the “Alternation” of French and English Leaders in the Liberal Party of Canada. Canadian Journal of Political Science 2 (1): 118-22.

- 22 - Reid, Escott. 1932. The Rise of National Parties in Canada. Papers and Proceedings of the Canadian Political Science Association. 4: 187-200. Saideman, Stephen M., David J. Lanoue, Michael Campenni, and Samuel Stanton. 2002. Democratization, Political Institutions, and Ethnic Conflict. Comparative Political Studies 35(1): 103-129. Sawatsky, John. 1991. Mulroney: The Politics of Ambition. Toronto: McClelland and Stewart. Smith, Denis. 1971. Bleeding Hearts … Bleeding Country: Canada and the Quebec Crisis. Edmonton: Hurtig. Ward, Norman. 1950. The Canadian House of Commons: Representation. Toronto: University of Toronto Press. Willison, Sir John S. 1903. Wilfrid Laurier and the Liberal Party: A Political History. Toronto: George N. Morang and Company. (2 vols.)

- 23 -

Appendices 1. Are the series stationary? For a series to be stationary, it must satisfy the “stability” condition and it must exhibit no trending, periodicity or structural breaks. For a series to be stable there must be a value on which it converges, in the absence of further shocks. The key is the value of d, the auto regressive component. If |d| < 1, the system converges. If |d| > 1, the system explodes, and the “equilibrium” is ± ¥, which in this paper is to say that Liberals either cover the entire landscape or simply disappear. If |d| = 1, then no equilibrium is identifiable; the system has a unit root and the trajectory is a random walk. The fact that each series in Table 2 has an autoregressive term less than one suggests that both are stable. The difficulty is that the conventional t test in this context underestimates the variance in the estimate, such that the system may seem convergent when it is not. To remedy this, I resort to Augmented Dickey-Fuller tests. As the boundary for convergence is a d with the absolute value of one, the null hypothesis for the test is that the system has a unit root. To make the null a coefficient with a value of zero, the test is performed on the coefficient for the shift from t -1 to t, rather than for the level at t. By definition, the lag term for the setup in changes is 1 - d. Results

Table A1. The stability condition

Interpolated Dickey-Fuller Critical Value MacKinnon p- value for Z(t) Test statistic 1% 5% 10% Quebec, 1867-2015 Z(t) -2.74 -3.64 -2.96 -2.61 0.07 Quebec, 1949-2015 Z(t) -1.88 -3.75 -3.00 -2.63 0.35 Rest of Canada, 1867-2015 Z(t) -4.38 -3.64 -2.96 -2.61 0.0003 Rest of Canada, 1949-2015 Z(t) -3.29 -3.75 -3.00 -2.63 0.02

of the test for each series are in Table A1.

1 For the rest of Canada, the picture is clear: the chance that the series is a random walk is vanishingly small. This is true for both the full series and the truncated one. For Quebec, however, we cannot definitively reject the hypothesis that the series is a random walk. The conventional t-test would suggest that the full series is not a random walk. The augmented values suggest that the chance is actually between a 5% and 10%. The MacKinnon combined estimate suggests that the series just misses the conventional threshold for rejection of the null. Nonetheless, I believe I can proceed, as the other indicators for stationarity send rather clearer signals. For the truncated Quebec series, the possibility seems real that the system is not stable. Estimates need to be treated with caution. On the matter of trend, the plots in Figures 1 and 2 suggest the following: • The Quebec series exhibits a positive trend to 1917, the conscription election, and a negative trend thereafter. Overall, a negative trend is discernible but barely so: a slope of Liberal share on consecutive observations of -0.24 but with a standard error, 0.19, almost as large as the slope. Awkwardly, the slope is steeper in the later years than in the earlier ones. • For the rest of Canada, the Liberal party colonized the landscape about a decade ahead of that in Quebec, and its share then rode a negative trend from virtually the start of the 20th century. Its downward drift is steadier than in Quebec, but is also shallower: a slope of -0.16 and a standard error of 0.09. Fortunately, the slope is shallow after World War Two. Visually, short-term departures from trend are at least as prominent as the long-term patterns. For neither series is there any suggestion of cyclicity. Parties win elections, often in sequence,

Table A2. Structural Breaks

Series Variable Statistic p-value Quebec LDV 11.01 0.02 Liberal leader 3.80 0.41 Rest of Canada LDV 3.83 0.40 Liberal leader 3.26 0.50 and then lose them, again often in sequence. But there is no suggestion of some underlying wave. Finally, there is the matter of structural breaks. In Table A2 I present “supremum” Wald tests for the Liberal leader variable and for the lagged dependent variable in each series. Although this is a weaker test than some alternatives, it represents the most comprehensive set of tests (Perron 2006). And none of the other tests suggests a fundamentally different result.

2 In three of the four situations, there is clearly no break in the series. The possible exception is for the lagged dependent variable in the Quebec series. A plot of test values does not suggest a significant break in 1917, where we might expect it to occur, but rather between 1972 and 1974. The attendant circumstance of this period do not obviously suggest an interpretation and a test for the equivalence of the LDV coefficient between the pre-1974 and post-1972 series does not indicate a significant difference. Substantively the value is 0.60 for the earlier estimation and 0.68 for the later one. Given the fact pattern, I have opted to use the full series. The supremum test is not calculable for the 1949-2015 series. 2. Can we estimate coefficients with OLS? One issue is whether serially correlated disturbances further bias our estimates of standard errors. This happens when residuals exhibit runs of consistently positive and negative values. Another issue is whether disturbances exhibit autoregressive conditional heteroscedasticity, that is, that their size (as distinct from their direction) varies systematically with time. Table A3 presents evidence from various tests. Although my use of the Koyck-lag scheme is motivated by theoretical considerations, the lag term can also soak up serial correlation in the disturbances. For

Table A3. Characteristics of disturbances, 1867-2015

DV lags Test c2 Pr > c2

Quebec Durbin 0.23 0.63 1 Breusch-Godfrey 0.27 0.60 ARCHLM 0.39 0.53 Rest of Canada Durbin 3.74 0.05 1 Breusch-Godfrey 3.86 0.05 ARCHLM 0.81 0.37 Durbin 0.91 0.34

2 Breusch-Godfrey 1.45 0.23 ARCHLM 0.25 0.61

this purpose, however, does the parsimonious setup with one lag suffice? For Quebec, it clearly does. By Durbin’s “alternative test,” the null hypotheses of no AR(1) correlation cannot be rejected. Similarly, by the Breusch-Godfrey test, higher-order correlations are also rejected. For

3 the rest of Canada, however, things are more complicated. By both tests, the single-lag setup appears to harbour serially correlated disturbances. Adding a second lag in the dependent variable seems to take care of the problem. The other worry, about heteroscedasticity, does not seem warranted. By the ARCHLM test, the null hypothesis of homoscedasticity is never rejected. The corresponding values for the 22-observation 1949-2015 series appear in Table A4. These are associated with the estimations in Tables 3 and 4. Hetereoscedasticity remains unproblematic for all estimations. Serial correlation is not a problem for the baseline estimation in Quebec but clearly is an issue for the estimation that also includes unemployment. For the rest of Canada, roughly the opposite is true. Durbin’s alternative measure signals first-order autocorrelation but gives reassurance for higher orders. Once unemployment is added, however, autocorrelation is washed out of the system.

Table A4. Characteristics of disturbances, 1949-2015

Baseline + Unemployment Quebec Durbin 2.42 (p = 0.12) 5.43 (p = 0.02) Breusch-Godfrey 2.89 (p = 0.09) 6.48 (p = 0.01) ARCHLM 1.53 (p = 0.22) 0.90 (p = 0.35)

Rest of Canada Durbin 10.79 (p = 0.01) 2.50 (p = 0.11) Breusch-Godfrey 8.64 (p = 0.09) 3.55 (p = 0.06) ARCHLM 1.53 (p = 0.22) 1.11 (p = 0.29)

Entries are c2 and associated p-values.

3. Influential Observations Are any single observations worrisomely influential? For example, one alternation that produces an exceptionally big shift will drag up the average. A basic test for the influence of individual observations is dfbeta, which shows how much and in what direction an excluded observation would pull the overall estimated effect, expressed in units of the estimated standard error of the coefficient. The conventional threshold for concern is an absolute dfbeta > 2 � = 0.309. The leader variable of special interest is the Liberal one.

4 Table A5. Influential Observations

Election Quebec Rest of Election Quebec Rest of Canada Canada

1867 | 1953 -0.03 0.04 1872 -0.04 1957 -0.12 -0.09 1874 0.04 -0.02 1958 0.12 0.19 1878 0.24 0.01 1962 0.08 -0.03 1882 0.09 0.06 1963 -0.08 -0.04 1887 -0.24 -0.28 1965 -0.24 0.06 1891 -0.16 0.06 1968 0.05 0.07 1896 0.02 -0.06 1972 -0.02 -0.16 1900 -0.04 0.28 1974 0.03 0.06 1904 -0.08 0.11 1979 0.06 -0.25 1908 -0.16 0.03 1980 0.06 -0.04 1911 -0.31 0.03 1984 0.20 0.19 1917 0.57 -0.09 1988 -0.03 -0.10 1921 -0.14 0.04 1993 -0.03 0.19 1925 0.02 0.06 1997 0.24 0.06 1926 -0.13 -0.07 2000 0.02 0.02 1930 0.03 -0.08 2004 -0.09 -0.05 1935 -0.10 -0.06 2006 -0.16 -0.14 1940 -0.21 -0.33 2008 -0.07 -0.33 1945 0.19 0.09 2011 0.36 0.22 1949 0.17 0.13 2015 0.10 0.43

Note: Entries are dfbetas for the impact of removing the observation from the estimation.

By this criterion, three elections might be worrisomely influential for Quebec. Two are the last elections for Laurier. In 1911, the party received its worst share since 1887 and in 1917, its best

5 ever. These are partly offsetting, although the absolute value for 1917 is roughly twice that for 1911. The other observation of note is for 2011. Here the combination is the absence of a Quebec leader and the party’s all time worst result. Three observations also stand out for the rest of Canada. The 1940 and 2008 elections pull the estimated effect down. In 1940, the conjunction is of a non-Quebec leader and an exceptionally strong result; this is the snap election called to give the incumbent Liberals a mandate for fighting World War Two without conscription for overseas service. The 2008 election conjoined a Quebec leader with the party’s worst result to date since 1867, worse even than 1917. Pulling the coefficient in the other direction is 2015, with a Quebec-based leader and a strong result following the 2011 low point. On balance, the outliers slightly weaken the rest-of-Canada result. 4. Equivalence of 1949-2015 Baseline Estimations to 1867-2015 Values Tests of the equivalence of baseline coefficients to those in the original estimations appear in Table A6. Without exception, the coefficients are statistically indistinguishable between the full- period and the truncated estimations. This is true for both Quebec and the rest of Canada.

Table A6. Equivalence of coefficients, truncated baseline to full-period

Quebec Rest of Canada

c2 p > c2 c2 p > c2

Liberal 0.90 0.34 2.22 0.14 Conservative 1.31 0.25 2.49 0.11 QC Ethnonational 0.11 0.74 Votes (t-1) 0.01 0.94 1.85 0.17 Votes (t-2) 0.01 0.93

5. Alternative Dynamic Specifications It is possible that the history in the system is better represented by lagging an independent variable rather than the dependent one. Although I see no particular theoretical warrant for any such move, the most plausible candidate is the Liberal leader variable. For Quebec, no setup suggests that lagging the term makes sense. In a setup in which both the leader variable and the dependent variable are lagged, impact from the LDV remains essentially undisturbed while estimated impact from the contemporaneous term of the IV is compromised. Lagging only the IV does not help matters. Where the lagged IV is entered along with the contemporaneous indicator, the former steals just enough impact from the latter to push it below

6 the conventional significance threshold even as the lag coefficient is only half the size of its standard error. If only the lagged term is entered, it captures some of the effect but only as a pale shadow, not much larger than its standard error. Both setups that omit the LDV have massively serially correlated errors, in contrast to the case with the LDV included. For Quebec, the case is clear: the system is autoregressive in the dependent variable. The same is true for the rest of Canada, but the story is more complicated. As I mention in the main text and in Appendix section 2, above, an LDV setup with one lag does not suffice for the rest of Canada. The starting point for comparing lags in the leader variable, then, is with the two- lag LDV setup. Adding a lag term for the Liberal leader variable leaves most of the architecture intact. The one-period lag on the vote is essentially as before. The two-period lag is eliminated, in effect, and it impact apparently absorbed by the leader terms. This setup leaves the error pattern undisturbed. The lagged leader effect is about the same size as the instantaneous one but with the opposite sign; the implication is that choosing a Quebec leader sets in train an oscillating effect in the rest of Canada. If we then drop the lag terms for the dependent variable, the oscillating structure in the leader terms remains as before. Statistics on the disturbances do not send warning signals, although each moves closer to the 0.05 threshold. If we drop the contemporaneous leader term, impact from the lag term remains negative but is appropriately smaller. The autocorrelation statistics all go into the danger zone: disturbances appear to be correlated at multiple lags and are conditionally heteroscedastic. What, if anything, should we make of the oscillating indication? In Figure 2, there is a suggestion of oscillation. The pattern is not confined to periods with Liberal leaders, although it may have regularized after 1960. In these years, Liberal leaders were Quebeckers more often than not. No mechanism comes to mind, apart from ordinary politics. And all these additional moves seem to be slicing the data more and more thinly. The relatively parsimonious model that was the starting point also seems to be the best place to stay.